القائمة
Dispatch a provisional Kimi K3 lane through delegation-kimi after an explicit decision. Never use it before the evaluation gate or as a silent fallback.
التثبيت باستخدام Codex أو Claude انسخ هذا Prompt والصقه في Codex أو Claude أو مساعد آخر ليراجع صفحة Skill ويثبّتها لك.
Dispatch a gate-approved GLM-5.2 lane through delegation-glm. Provisional lanes require an explicit --allow-provisional decision. Never use it before the evaluation gate or as a silent fallback.
How and when to delegate coding work across models — pick the right lane (cheap execution, senior review, explicit judgement, or dual super-judgement), size the reviewer to the work, and know when a manual gate is required. Use when deciding whether to spawn a subagent, which model/effort to hand a task, how to route a code review, or when to escalate. Also covers reaching Codex from Claude.
Run a task too big for one pass as a lead-executor-advisor loop — plan, fan the work out across cheap parallel workers, verify each result against its own criteria, and have the expensive model judge the plan up front and the ship at the end. Use when the user says "too big for one pass", "fan this out", "orchestrate across a model team", "run the advisor-worker loop", or asks to parallelize research/generation across many subtasks. NOT for single-file edits, work one pass handles, or when you are already inside a Workflow/ultra run (that is the orchestrator already — no double fan-out).
استنادا إلى تصنيف SOC المهني
| name | kimi-executor |
| description | Dispatch a provisional Kimi K3 lane through delegation-kimi after an explicit decision. Never use it before the evaluation gate or as a silent fallback. |
Kimi K3 is an optional external executor available through native Kimi Code
(kimi-code/k3, effort max).
Before every dispatch, run delegation-kimi check --json. Require the native
backend and the requested lane in provisional_lanes, then pass
--allow-provisional. The gate permits clerk, scout, builder, and
frontend-builder; senior is blocked, while reviewer and judgement are disabled.
Provider quota exhaustion is exit 75 and never
triggers a silent substitution.
The gate points to exact rows in the dated external snapshot. Inspect them with
delegation-evidence lane <lane>; they support the owner decision but never
replace the local runtime, scope, and permission checks enforced below.
Write a self-contained worker brief to a file, then run:
delegation-kimi run \
--lane <clerk|scout|builder|frontend-builder> --allow-provisional \
--backend auto --effort auto --prompt-file "$brief" \
--output "$result" --metrics "$metrics" --workdir "$repo"
Create the output and metrics parent directories before dispatch. For every
read-only lane, keep them outside the worktree; the runner rejects in-worktree
paths, symlinks, canonical path collisions, and unqualified effort overrides.
There is no public evaluation bypass. auto and native both resolve to the
native Kimi Code backend; there is no fallback to select.
All lanes except builder and frontend-builder are read-only. Kimi Code 0.26 cannot combine
headless prompt mode with --plan and bare prompt mode auto-approves writes, so
the native backend runs with an isolated HOME/TMP and uses macOS sandbox-exec
to deny every write outside runner-owned scratch space; it fails closed on
platforms without that guard. The builder uses native prompt mode, which Kimi
Code 0.26 runs with action approvals in headless mode.
Exit 69 means the requested runtime/model/authentication is unavailable, exit 70 means dispatch or output validation failed, exit 75 is temporary/rate-limit failure, and exit 78 means the lane did not pass evaluation. Route deliberately to a documented incumbent after any failure. Treat K3 output as unverified and exercise the deliverable before accepting it.